OBJECTIVES

Effect modification (EM) causes bias in network meta-analyses (NMA) if EM varies across treatments in the network. Several methods have been developed, dealing with EM in NMAs when aggregated data (AgD) is available for at least one trial in the network. These methods typically make the shared effect modification assumption (SEMA) and disregard the available information the EM contains on the relative treatment effect (RTE) in the AgD trials. The SEMA is debatable, especially when comparing different classes of therapies. Our aim is to present a novel NMA technique that considers all available information on the impact of EMs on the RTE in the evidence network.

METHODS

We simulated an evidence network of seven trials: three A-B trials, three A-C trials (with both sets reported as AgD) and one A-D trial reported as individual patient data (IPD), using a binary outcome and two correlated EMs, following SEMA within but not across comparisons. The novel network meta-interpolation (NMI) technique combines the intention-to-treat RTE with typically reported subgroup RTE to predict the RTE and its uncertainty at a standardized combination of the two considered EMs for all trials. The IPD trial informed missing EM values. NMI results at standardized EM values were then compared with standard NMA and network meta-regression (NMR) over 1000 random simulations.

RESULTS

Averaging over all RTEs, NMI achieved an average mean squared error of 0.045 relative to 0.106 by NMA and 0.196 by NMR and 95% credible interval coverage of 92.4% relative to 70.8% and 68.2%, respectively.

CONCLUSIONS

The NMI is a novel method not relying on the SEMA, convincingly outperforming NMA and NMR in this simulation study. This approach can rely fully on AgD in case of one EM and requires one IPD trial in case of multiple correlated EMs. Comparison against multilevel-NMR and real applications are forthcoming.